Automated Segmentation of the Cervix in Transvaginal Ultrasound Images

Preterm birth (PTB) is a major global public health concern, affecting approximately 15 million pregnancies annually. One of the most recognized biomarkers for PTB risk assessment is cervical length (CL), which is typically measured manually from transvaginal ultrasound (TVUS) images of the cervix. However, this manual process is prone to inter-operator variability and subjectivity, limiting its reliability of early risk assessment. Standardizing the segmentation of a region of interest in the cervix could enhance the consistency of CL measurements and improve the detection of high-risk pregnancies. This study proposes a U-Net convolutional neural network for segmenting the mid-third of the cervix in TVUS images obtained during the first trimester of pregnancy. 224 images were collected from 249 pregnant women from two specialized maternal-fetal health centers. To evaluate the model, we employed the Dice coefficient, mean surface distance (MSD), and area error rate (AER). The results show that the model achieved a Dice coefficient of 0.84 (95% CI: 0.78-0.88), MSD of 8.89 mm (95% CI: 5.47-12.31), and AER of 607 mm 2 (95% CI: 402–812). This work represents a contribution towards fully automated, computerized TVUS image analysis for PTB risk assessment.